Wan

Wan is a family of open-weight video models from Alibaba, spanning both text-to-video (T2V) and image-to-video (I2V). The line is built around Alibaba's Wan-VAE, an efficient video autoencoder, and Wan 2.1 was notable as the first open video model able to render both Chinese and English text inside a clip. It scales across sizes — from a lightweight 1.3B T2V model that fits in about 8GB of VRAM up to the flagship 14B models — so the same ecosystem covers everything from quick consumer-GPU drafts to high-fidelity cinematic generation.

The versions diverge in architecture and focus. Wan 2.2 introduces a Mixture-of-Experts (MoE) design that splits the denoising process across specialized expert models, enlarging capacity at the same compute cost, and adds a fast 5B TI2V variant (16×16×4 compression via Wan2.2-VAE) that generates 720p at 24fps on a card like a 4090; its A14B T2V and I2V models were trained on substantially more data with curated cinematic-aesthetic labels for lighting, composition, and color. Wan 2.5 is the hosted, audio-aware generation — synchronized voices, ambient sound, and music alongside cinematic 1080p output — while Wan 2.7 pushes temporal consistency and subject stability, reducing flicker, distortion, and identity drift across frames.

Choose Wan when you want open, controllable video with strong native image-to-video and the deepest LoRA and motion-effect ecosystem in open video generation. Reach for the 2.2 A14B models as the open-weight workhorses for T2V and I2V, the 5B variant when speed matters, and the hosted 2.5 / 2.7 releases when you want the best prompt adherence, motion realism, or audio-synced clips. Against closed APIs like Kling you trade some polish for open weights, stackable LoRAs, and no local install when you run it here.